MCP Technical Glossary

Local process communication over standard input (stdin) and standard output (stdout).

The primary transport mechanism for local MCP servers. The host application spawns the MCP server as a subprocess and exchanges JSON-RPC 2.0 frames over stdin and stdout pipes. Important rule: STDOUT is strictly reserved for JSON-RPC messages; all debug logging must go to STDERR to prevent parse errors.

HTTP-based transport using unidirectional streaming from server to client with POST endpoints.

A remote transport mechanism enabling AI clients to connect to MCP servers hosted on remote servers, Docker containers, or cloud endpoints. The client opens an HTTP GET connection receiving a text/event-stream of JSON-RPC notifications and dispatches requests via HTTP POST.

Modern standard HTTP streaming transport for stateless remote microservices.

Introduced in the mid-2026 MCP specification update. Allows remote MCP servers to operate as standard stateless HTTP workloads without requiring persistent session handshakes, making horizontal scaling trivial.

Executable functions exposed by servers that AI models can autonomously invoke.

Tools are dynamic functions registered by MCP servers. Each tool specifies a name, description, and JSON Schema input parameter definition. The host LLM decides when to invoke a tool, and the client sends a `tools/call` request to execute it and return the result to the conversation context.

Read-only context data attached to conversations using standard URI schemes.

Resources provide contextual information (such as database schemas, log files, documentation, or git diffs) to AI models. Resources are identified by custom URIs (e.g. `postgres://tables`, `file:///repo/README.md`) and can be subscribed to for real-time update notifications.

Pre-engineered slash command templates with user parameter inputs.

Prompts allow MCP servers to publish interactive prompt templates into AI clients (like Claude Desktop or Cursor). Users can invoke these templates directly in chat as slash commands (e.g. `/review-pr` or `/explain-code`) with typed arguments.

Workspace filesystem boundaries communicated from client to server.

Roots define the directory boundaries of the active user project. The host client notifies connected servers of allowed workspace folders, ensuring servers like filesystem or git operate strictly within permitted directories.

Allows an MCP server to request LLM completions back through the host client.

A two-way mechanism where a running MCP server can ask the host AI client to generate an LLM completion. This enables servers to perform autonomous sub-agent loops, recursive summarization, or semantic classification without needing their own API keys.

The application initiating and orchestrating the MCP connections (e.g. Claude Desktop, Cursor).

The Host is the user-facing AI application or agent runtime. It manages the lifecycle of MCP servers, translates model intentions into JSON-RPC tool calls, verifies user permissions, and injects server results into the prompt context.

Stateless, lightweight remote procedure call protocol formatted in JSON.

The foundational message encoding standard used by Model Context Protocol. All requests, responses, notifications, and error objects conform to the standard JSON-RPC 2.0 specification with `jsonrpc: '2.0'`, `id`, `method`, and `params` fields.

Printing non-JSON log messages to stdout, crashing stdio transport parsing.

A common bug where server code uses `console.log()` or `print()` to output debug strings. In stdio transport, this corrupts the JSON-RPC stream, causing JSON parse errors (-32700) or internal error (-32603) in the host client. Always direct debug logs to STDERR.